Hugging Face's logo

Datasets:
Barthel
/
variouscryptodata

Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
Size:
100M - 1B
Tags:
finance
prediction-markets
polymarket
order-book
market-microstructure
crypto
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
variouscryptodata / README.md
Barthel's picture
Barthel
dataset card: hyperliquid_mainnet_archive part 2 (dedicated collector), bbo_stream config, audited corrections
672ba05 verified
preview code
|
Raw
History Blame Contribute Delete
13.6 kB
metadata 
license: cc-by-4.0
pretty_name: Various crypto market data (Polymarket Up/Down order books, …)
language:
  - en
tags:
  - finance
  - prediction-markets
  - polymarket
  - order-book
  - market-microstructure
  - crypto
  - time-series
size_categories:
  - 100M<n<1B
configs:
  - config_name: polymarket_updown_orderbook
    data_files: polymarket_updown_orderbook/data/*/*.parquet
    default: true
  - config_name: hyperliquid_trades
    data_files: hyperliquid_trades/data/*/*.parquet
  - config_name: hl_archive_l2book
    data_files: hyperliquid_mainnet_archive/l2book/*/*.parquet
  - config_name: hl_archive_trades
    data_files: hyperliquid_mainnet_archive/trades/*/*.parquet
  - config_name: hl_archive_asset_ctx
    data_files: hyperliquid_mainnet_archive/asset_ctx/*/*.parquet
  - config_name: hl_archive_bbo
    data_files: hyperliquid_mainnet_archive/bbo/*/*.parquet
  - config_name: hl_archive_mark
    data_files: hyperliquid_mainnet_archive/mark/*/*.parquet
  - config_name: hl_archive_funding
    data_files: hyperliquid_mainnet_archive/funding/*/*.parquet
  - config_name: hl_archive_candles_5m
    data_files: hyperliquid_mainnet_archive/candles/*/*.parquet
  - config_name: hl_archive_bbo_stream
    data_files: hyperliquid_mainnet_archive/bbo_stream/*/*.parquet

variouscryptodata

Crypto market datasets collected as a by-product of our own research and published so they are not lost. One sub-folder per dataset; each appended nightly where collection is still running.

folder what coverage cadence
polymarket_updown_orderbook/ Polymarket Up/Down (5m/15m) order books, 10 levels, BTC/ETH/SOL/XRP/DOGE/HYPE/BNB, with Binance spot reference 2026-05-24 → present appended nightly (previous UTC day)
hyperliquid_trades/ Hyperliquid perp trade prints (every fill) for BTC/ETH/SOL/HYPE/AVAX + 25 HIP-3 tradfi markets (xyz:* equities/commodities/FX, cash:*, km:*) 2026-07-23 → present appended nightly (previous UTC day)
hyperliquid_mainnet_archive/ Hyperliquid WebSocket capture for 16 perps (15 until 2026-07-20): L2 order book snapshots (20 levels/side), trades, asset context (funding/OI/premium/oracle/mark/mid/impact), top-of-book+depth summary, mark/oracle, funding, 5m candles part 1: 2026-07-18 (14:03 UTC) → 2026-08-04 (13:25 UTC), 16 perps (legacy bot); part 2: from 2026-08-22 ~19:49 UTC (first day partial), all ~320 perp + HIP-3 markets (dedicated collector) part 2 appended nightly (previous UTC day)

polymarket_updown_orderbook

Continuous top-10-level order book snapshots for Polymarket's short-dated crypto Up/Down markets (5-minute and 15-minute, BTC/ETH/SOL/XRP/DOGE/HYPE/BNB), captured every ~6 seconds per market (≈2–3 s in the last minute before a market closes), with the Binance spot price, the market's strike and the time left to resolution on every row.

  • Granularity: one row per (market, snapshot); ~250 000 rows/day.
  • Both outcomes: the Up and Down books are recorded side by side.
  • Layout: polymarket_updown_orderbook/data/date=YYYY-MM-DD/book_depth.parquet (Hive-style day partitions).

Polymarket's CLOB API is live-only; to our knowledge no public archive of the order books of these short-dated markets exists.

Schema

column type meaning
ts int64 snapshot time, Unix seconds (UTC)
ts_ms float64 snapshot time, Unix milliseconds (v2 rows; absent before 2026-06-12)
market string market slug, e.g. btc-updown-5m-1779622200 (asset-period-windowStart)
asset string btc, eth, sol, xrp, doge, hype, bnb (all seven present on every day)
period_min int64 market length in minutes: 5 or 15
condition_id string Polymarket condition id (public market identifier, 0x…)
win_start_ts int64 window start, Unix seconds (v2)
secs_left int64 seconds until the window closes / market resolves
in_window bool true only inside the collector's own entry window, 8–35 s before the market closes (not an 'is the market open' flag — every market is captured for its whole 5/15-minute life)
favored string which side has the higher mid-price (Up/Down); null when tied or when either book lacks a two-sided quote (common in the final minute; ~15–25 % of rows)
spot float64 Binance spot price of the underlying (REST ticker) fetched with the snapshot
spot_ts_ms float64 timestamp of that spot observation, Unix ms (v2)
strike_spot float64 first spot at/after window start = the market's strike (v2)
Up string (JSON) {"bids":[[price,size],…],"asks":[[price,size],…]} — up to 10 levels each, best first
Down string (JSON) same for the Down outcome

Prices in Up/Down are outcome-token prices in USDC (0–1); sizes are in shares. Parse with json.loads (Python) or json_extract (DuckDB). Schema v1 (19 days, 2026-05-24 → 2026-06-11) has 11 columns; v2 (from 2026-06-12) has 15. Reading with union_by_name/diagonal concat handles both.

Quick start 

import polars as pl, json
df = pl.read_parquet(
    "hf://datasets/Barthel/variouscryptodata/polymarket_updown_orderbook/data/date=2026-08-21/book_depth.parquet")
row = df.filter(pl.col("asset") == "btc").row(0, named=True)
book = json.loads(row["Up"])
print(book["bids"][0], book["asks"][0], row["spot"], row["secs_left"])

-- DuckDB
SELECT asset, period_min, count(*)
FROM read_parquet('hf://datasets/Barthel/variouscryptodata/polymarket_updown_orderbook/data/*/*.parquet', union_by_name=true)
GROUP BY 1,2 ORDER BY 1,2;

Collection notes (read before modelling)

  • Best-effort single-host capture: short gaps (seconds to minutes) occur around reconnects and host maintenance; treat ts spacing as irregular.
  • Snapshot cadence is adaptive: ~6 s per market normally, ~2–3 s during the last minute before a market closes; all live markets (7 assets × 2 periods) are polled in the same cycle.
  • spot is the Binance spot price fetched by the collector at snapshot time, not Polymarket's resolution oracle; use it for analysis, not as ground truth for settlement.
  • Nothing here is investment advice; no trading strategy is included.
  • Every upload is scanned automatically for credentials before publishing.

hyperliquid_trades

Every public trade print streamed from Hyperliquid's WebSocket trades channel for 30 markets: the perps BTC, ETH, SOL, HYPE, AVAX and 25 HIP-3 markets (tokenised equities, indices, commodities and FX such as xyz:NVDA, xyz:TSLA, xyz:GOLD, xyz:SP500, xyz:EUR, cash:USA500, km:US500 — the coin column uses Hyperliquid's dex:NAME form). One Parquet per UTC day, all markets in one file. Roughly 100 rows per day carry exchange timestamps far outside the file's day: on every (re)subscription Hyperliquid replays a market's most recent trades, and for the three dormant markets cash:SILVER, cash:USA500, km:US500 (no trades at all during the collection period so far) those ~30 replayed trades date from June/July 2026 and recur in every daily file; filter on time_ms if that matters to you.

  • Layout: hyperliquid_trades/data/date=YYYY-MM-DD/trades.parquet
  • Rows: ~0.6–3.6 million/day (median ≈2 M; weekends lowest); sorted by time_ms; de-duplicated on (coin,tid).
column type meaning
coin string Hyperliquid market name (BTC, xyz:NVDA, …)
side string aggressor side as reported by Hyperliquid: B = buyer, A = seller
px float64 trade price (USDC)
sz float64 trade size (base units of the market)
time_ms int64 exchange trade time, Unix milliseconds (UTC)
tid int64 Hyperliquid trade id

Deliberately not included: counterparty wallet addresses (users) and transaction hashes — this dataset is about prices and flow, not about who traded. Collection is best-effort from a single WebSocket client; brief gaps (reconnects) can occur. HIP-3 markets follow their own trading hours, so zero-trade stretches there are normal, not gaps.

import polars as pl
t = pl.read_parquet("hf://datasets/Barthel/variouscryptodata/hyperliquid_trades/data/date=2026-08-21/trades.parquet")
print(t.group_by("coin").agg(pl.len(), (pl.col("px")*pl.col("sz")).sum().alias("notional")).sort("notional", descending=True).head(10))

hyperliquid_mainnet_archive (part 1 static 2026-07-18 → 2026-08-04; part 2 from 2026-08-22, appended nightly)

An 18-day capture (2026-07-18 ~14:03 UTC → 2026-08-04 ~13:25 UTC; first and last day partial) of Hyperliquid's public WebSocket feed by a shadow-trading research bot (no orders were sent from this data). Coins were the bot's watch-list at the time — 16 liquid perps: BTC, ETH, SOL, HYPE, XRP, AVAX, NEAR, ONDO, UNI, WLD, ZEC, PUMP, TRUMP, FARTCOIN, LIT, VVV (15 coins on 2026-07-18 → 07-20; AVAX was added 2026-07-21, 16 from then on). One Parquet per table per UTC day: hyperliquid_mainnet_archive/<table>/date=YYYY-MM-DD/<table>.parquet.

table rows/day (≈) columns
l2book 100–300 k coin, time_ms (exchange), recv_ms (local receive), bids, asks — JSON [[px, sz, n_orders], …], 20 levels per side in part 1 (up to 20 in part 2), best first, full precision (nSigFigs=null); one snapshot per coin every ~5.4 s (the exchange's push cadence, not every book update)
trades 0.3–1.3 M coin, side (B buyer-aggressor / A seller-aggressor), px, sz, time_ms, tid — de-duplicated on (coin,tid); wallet addresses and tx hashes removed
asset_ctx 0.5–1.5 M coin, recv_ms, funding (hourly rate), open_interest, prev_day_px, day_ntl_vlm, day_base_vlm, premium, oracle_px, mark_px, mid_px, impact_bid, impact_ask — streamed activeAssetCtx updates
bbo 100–300 k coin, recv_ms, bid_px, bid_sz, ask_px, ask_sz, spread, spread_bps, bid_depth_sz, bid_depth_usd, ask_depth_sz, ask_depth_usd, bid_levels, ask_levels — top of book plus summed depth over the 20 captured levels, computed by the collector for each captured l2book snapshot (same row count and ~5 s cadence as l2book)
mark ~250 k coin, recv_ms, mark_px, oracle_px
funding ~22 k coin, recv_ms, funding_rate
candles 2–5 k coin, interval (5m), open_ms, close_ms, open, high, low, close, volume, trade_count — final state of each 5m candle
bbo_stream (part 2 only) ~5–10 M coin, time_ms, recv_ms, bid_px, bid_sz, bid_n, ask_px, ask_sz, ask_n — Hyperliquid's high-frequency bbo WebSocket channel (every top-of-book change) for the ~40 highest-volume markets

Part 2 (from 2026-08-22 ~19:49 UTC, first day partial; appended nightly): a dedicated read-only collector subscribes l2Book + trades for every perp on the main exchange and every HIP-3 market (≈320 markets, re-discovered every 6 h, so newly listed markets appear automatically), and bbo for the ~40 highest-volume markets (ranked at collector start; new high-volume entrants are added at discovery, none removed). Same table layout and column names as part 1, with these differences:

  • asset_ctx, mark and funding come from the REST metaAndAssetCtxs endpoint once per minute (part 1: streamed, ~5 s).
  • bbo is derived from each l2book snapshot (identical definition to part 1).
  • l2book: up to 20 levels per side — thin HIP-3 markets can have fewer (see bid_levels/ask_levels); the ~5.4 s cadence is the exchange's push rate.
  • candles are 5-minute bars built from trades captured live (receive latency ≤ 60 s), only buckets with ≥ 1 trade (part 1: exchange candle stream incl. empty buckets); buckets around collector (re)starts may be partial.
  • trades: on every (re)subscription Hyperliquid replays the last ~30 trades of a market, so each daily file can contain a few older trades per market — filter on time_ms if that matters.
  • Part-2 volumes: l2book/bbo ≈ 5 M rows/day, trades ≈ 5–8 M, bbo_stream ≈ 9–11 M, asset_ctx/mark/funding ≈ 0.46 M, candles ≤ 92 k.
  • Receive latency (recv_ms − time_ms): l2book ≈ 0.6 s median / ≈ 1–2 s p99; trades/bbo ≈ 0.35 s median; bursts up to ~10 s at (re)subscribe.

Gap between part 1 and part 2: 2026-08-04 13:25 → 2026-08-22 19:49 UTC.

Notes: time_ms/open_ms are exchange timestamps; recv_ms is the collector's receive time (single host, best-effort; latency typically ≈0.45 s median and ≈1 s p99, with occasional bursts up to ~10–30 s). Hyperliquid's own complete history is available from the exchange's requester-pays S3 archive; this is a free, partial mirror for convenience.

import polars as pl, json
b = pl.read_parquet("hf://datasets/Barthel/variouscryptodata/hyperliquid_mainnet_archive/l2book/date=2026-07-27/l2book.parquet")
snap = b.filter(pl.col("coin") == "ETH").row(0, named=True)
bids, asks = json.loads(snap["bids"]), json.loads(snap["asks"])
print(bids[0], asks[0])   # [px, sz, n_orders]

License & citation

Data © the collector, released under CC-BY-4.0. Underlying quotes originate from Polymarket's public CLOB API, Binance's public REST API and Hyperliquid's public WebSocket API. If you use this data, please cite “Barthel/variouscryptodata (Hugging Face dataset)”.